200
Best Efforts level of service. However, there is room for introduction of a new Class 5, if it is required. I use this value later when I talk about Expedited Forwarding.
This set of definitions for the AF DSCP is clearly compatible with the older TOS format. The only difference is that the older definitions of delay, throughput, and reliability are replaced with a new two-bit
pattern indicating drop precedence. The last bit is always equal to zero.
There is only one defined DSCP value for EF. RFC 2598 recommends using the value 101110 for this purpose. Note that this is the obvious extension to the values in
Table 8-2 . Since EF offers service
guarantees that are not available in AF, it is in some sense a higher priority. One additional Class value is available before reaching the reserved values—the value 101, which would be Class 5. At the same time,
since packets designated for EF should not be dropped, they have the highest drop precedence value, 110. This value inherently means that only one type of EF is available.
A network cant have, for example, two flavors of EF—one with low and the other with high reserved bandwidth. If this separation is required, the best strategy is to define additional Control Point values and
configure the routers to recognize them. In this case, it is better to fix the first three bits at 110 and use the second three bits to specify the different forwarding characteristics. However, it is important to remember
that devices on other networks such as the public Internet will not recognize these parameters and may not handle the packet as delicately as you would like.
8.6.6 Traffic Shaping
Traffic Shaping is a system for controlling the rate of data flow into a network. Networks often use it in conjunction with other QoS mechanisms.
There are two main ways to control the rate of flow of traffic. A device can either throw away packets whenever the specified rate limit is reached, or it can buffer packets and release them at the specified rate.
The process of discarding packets that exceed a bandwidth parameter is usually called policing. Saving packets for future transmission is called buffering.
In any real-world application, of course, it is necessary to do both policing and buffering. If an application persistently sends data at twice the rate that the network can forward it, then it doesnt matter how many
packets are put into the buffer because the network simply cant send them all along. If the traffic flow is characterized by a number of short bursts, then network devices can easily buffer the bursts to smooth
them out—provided, of course, that the time average of the traffic rate is less than the available output bandwidth.
Traffic Shaping can be done either on an entire pipe of incoming data or on individual data flows. Usually, network designers implement Traffic Shaping only at network bottlenecks and at input points into the
network.
EF is a good example of a place where Traffic Shaping needs to be used in conjunction with a QoS mechanism. The EF model specifies a certain sustained bandwidth level that the data flow is allowed to
use. If an application exceeds this flow rate, then the excess packets are dropped. The best way to implement such a service is to ensure that the data stream entering the network is restricted to less than the
reserved bandwidth. This data flow may enter the network from a user segment within the network, in which case the first router the traffic encounters does the traffic shaping.
Dropping packets, while undesirable, is not a completely bad thing in a network. Many protocols such as TCP have the ability to notice when they start losing packets because every packet has a sequence number.
If packets do not follow in sequence then the end devices usually wait a short time to see if the missing packet will eventually arrive. When this time has elapsed, the receiving device sends a notification to the
sending device to tell it about the missing packet.
201
When this happens, the sender assumes that the network has a reliability problem, and it reduces the number of packets it will send before it gets an acknowledgement the TCP Window. Reducing the
packets also reduces the amount of bandwidth that the application consumes.
By dropping TCP packets, the network can effectively control the rate that the application sends data. It tends to back off until it no longer sees dropped packets. This data-flow rate is exactly equal to the preset
Traffic Shaping limit.
However, not all applications behave as well as TCP when they suffer from dropped packets. For example, UDP packets generally do not need to be acknowledged. Thus, UDP applications may not respond
properly to traffic shaping. IPX has similar issues. The connection-oriented SPX protocol can respond to dropped packets by reducing its windowing, similar to TCP. But other IPX protocols are not so well
behaved.
In general, it is a good idea to monitor applications that use heavily policed links to ensure that they behave well. If they do not, then you must increase the bandwidth to reduce the congestion.
8.6.7 Defining Traffic Types
